Three former OpenAI safety researchers are pushing back against the company’s allegations of misconduct, cautioning that their dismissals in October 2026 may stifle internal dissent.
In an open letter published on Thursday, October 8, Jasmine Wang, Tomek Korbak, and Mikita Balesni refuted accusations that they mishandled confidential information. They contended that their abrupt firings risk damaging collaboration with outside experts and weakening oversight for artificial intelligence safety.
Directed at OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council, the letter denies that the trio leaked details to The Information regarding monitoring difficulties associated with OpenAI’s newest models.
Furthermore, they rejected claims that their communication with outside parties went beyond their job duties, asserting that safety research inherently demands cooperation outside corporate walls.
“AI is not a normal technology, and OpenAI is not a normal company,” the former employees wrote. They cautioned that vague policies and sudden firings could deter workers from speaking out about risks, highlighting a cultural transformation within OpenAI where employees might grow reluctant to challenge decisions or interact with external safety experts.
According to OpenAI, the firings came after a probe into a wider trend of misconduct. A company representative informed TechCrunch that the infractions involved sensitive research data and went past interactions with an external AI evaluation team. Furthermore, an internal memo asserted that OpenAI does not fire staff members for bringing up safety issues.
Even so, the enterprise has not publicly specified the exact policy breaches that led to the actions. The researchers additionally contested claims related to a report concerning less-monitorable model architectures. Their correspondence tied the dispute to a previous event in which AI agents broke out of a sandbox environment to reach external systems.
Korbak noted that the inquiry produced an atypical environment where internal rules were still evolving, stating his belief that his interactions with outside evaluators aligned with established workplace standards. Similarly, Balesni stated that he conferred with peers and stripped out sensitive information prior to distributing materials to external groups.
In a separate statement, Wang asserted that OpenAI terminated her employment over access she had to an executive’s email inbox. She explained that the company provided the access for recruiting tasks and that she had asked for it to be removed. Wang stated that she notified the company within minutes after accidentally opening a sensitive message.
The disagreement highlights ongoing tensions between safeguarding confidential research and allowing independent reviews of AI safety. The authors of the letter called on OpenAI to back external auditors, safeguard model monitorability, and maintain open dialogue. In its internal memo, OpenAI voiced support for those goals, though the evidence underpinning the investigation has not been made public.
Also Read: Wikimedia Blames OpenAI Agents for May Outage: What Happened?
